I started working on Aether after watching my mother fall sick.
She was diagnosed with an exceptionally rare form of cancer only a few days before her death. At the time, fewer than ten people in the world had ever been diagnosed with it. By the time the diagnosis was made, there was no meaningful window left for intervention.
Her medical records existed. Every test, every report, every prescription was there somewhere. But no one ever had a complete picture of her health over time. Each specialist saw a slice. Each visit reset context. Signals that might have mattered only became obvious in retrospect.
This was not negligence. It was how the system was structured.
That experience made one thing painfully clear to me: medicine is longitudinal, but our systems are episodic.
Why rare diseases reveal how episodic medicine breaks down
Rare diseases are often described as edge cases. Individually uncommon, statistically inconvenient, hard to study. In reality, they expose something much more fundamental about modern medicine: our inability to reason across time.
As the authors of a recent npj Digital Medicine study note:
They add that patients experience a prolonged diagnostic odyssey, often involving multiple clinicians and years of delay before an accurate diagnosis is reached.
An estimated 300 to 400 million people worldwide live with a rare disease. Most of them do not lack data. They lack continuity. Their medical history is scattered across hospitals, labs, specialists, PDFs, portals, and memories.
The problem is not rare disease complexity alone. The problem is how we record, reason about, and learn from health data over time.
Why labels are the wrong abstraction for medicine
Much of healthcare AI assumes that diagnosis labels represent ground truth. In practice, they rarely do.
The paper states this directly:
Yet most machine learning systems treat these labels as equally true and equally final.
Rather than assuming clean labels, the study starts from a more honest premise:
This framing mirrors real clinical practice. Diagnoses are often probabilistic. They evolve. They get revised. Systems that force binary certainty too early inevitably miss slow, atypical, or rare conditions.
What the research actually shows
The study introduces a weakly supervised transformer framework that combines a small number of expert confirmed cases with a much larger pool of probabilistic labels derived from real world EHR data. These weaker labels are refined iteratively as the model learns.
As the authors describe it:
When evaluated on rare pulmonary diseases, the approach outperformed rule based methods, gradient boosted models, and standard transformer architectures trained only on expert labels.
But the most important finding was not diagnostic accuracy alone.
The model uncovered clinically meaningful subphenotypes.
“The learned patient representations revealed latent structure associated with clinically meaningful heterogeneity, including subgroups with significantly different progression and survival.”
Patients with the same diagnosis showed fundamentally different disease trajectories. These differences were not predefined. They emerged from longitudinal patterns in the data.
The lesson is not about transformers
It would be easy to read this paper as another example of transformers outperforming older models. That misses the point.
The deeper insight is about representation.
Instead of optimising for a single endpoint, the model learns what the authors describe as:
Once such a representation exists, diagnosis becomes just one of many questions that can be asked. Risk stratification, progression modelling, and subtype discovery become natural extensions rather than separate systems.
This is what longitudinal health requires.
Why this matters for how we build health systems
Most EHRs are designed to document encounters, not to learn from them.
The study makes this limitation explicit:
Rare diseases simply make this failure visible sooner.
If disease understanding evolves over time, then health systems must preserve uncertainty, allow diagnoses to be revised, and connect clinical events across years rather than visits.
This is not a call for more features. It is a call for better foundations.
Aether’s perspective on longitudinal health
At Aether, we think of health not as a collection of reports, but as a continuously evolving state.
The health graph we are building is designed to reflect what this research validates:
“Medicine is longitudinal, but much of health AI remains episodic.”
Our goal is not perfect documentation. It is continuity. An EHR Lite approach is not about replacing hospital systems. It is about restoring longitudinal coherence where it is most often lost, in outpatient care, diagnostics, referrals, and chronic conditions that unfold over years.
Research like this reinforces the direction we have taken. Learning from weak, imperfect, real world data is not a compromise. It is the only path that aligns with how medicine actually works.
From documentation systems to learning systems
The future of medicine is not better labels. It is better learning over time.
Some diseases are only understood very late. Others only in retrospect. Our systems should not make that inevitability worse by fragmenting context and freezing assumptions.
Rare disease AI is not just about rare diseases. It is a mirror held up to the rest of medicine.
And it is quietly showing us what longitudinal health should look like.
References
Greco, K.F., Yang, Z., Li, M. et al. A weakly supervised transformer for rare disease diagnosis and subphenotyping from EHRs with pulmonary case studies. npj Digital Medicine (2026).
https://www.nature.com/articles/s41746-026-02406-x
https://www.nature.com/articles/s41746-026-02406-x_reference.pdf